On 12/10/2014 08:02 AM, Eric Dumazet wrote:

On Tue, 2014-12-09 at 19:40 -0800, Alexander Duyck wrote:

quoted

I also took the opportunity to refactor the core bits that were placed in
__alloc_page_frag.  First I updated the allocation to do either a 32K
allocation or an order 0 page.  This is based on the changes in commmit
d9b2938aa where it was found that latencies could be reduced in case of
failures. 

GFP_KERNEL and GFP_ATOMIC allocation constraints are quite different.

I have no idea how expensive it is to attempt order-3, order-2, order-1
allocations with GFP_ATOMIC.

The most likely case is the successful first allocation so I didn't see
much point in trying to optimize for the failure cases.  I personally
prefer to see a fast failure rather than one that is dragged out over
several failed allocation attempts.  In addition I can get away with
several optimization tricks that I cannot with the loop.

I did an interesting experiment on mlx4 driver, allocating the pages
needed to store the fragments, using a small layer before the
alloc_page() that is normally used :

- Attempt order-9 allocations, and use split_page() to give the
individual pages.

Boost in performance is 10% on TCP bulk receive, because of less TLB
misses.

With huge amount of memory these days, alloc_page() tend to give pages
spread all over memory, with poor TLB locality.

With this strategy, a 1024 RX ring is backed by 2 huge pages only.

That is an interesting idea.  I wonder if there would be a similar
benefit for small packets.  If nothing else I might try a few
experiments with ixgbe to see if I can take advantage of something similar.

- Alex